U.S. patent application number 11/996364 was filed with the patent office on 2008-09-04 for audio signal modification.
This patent application is currently assigned to KONINKLIJKE PHILIPS ELECTRONICS, N.V.. Invention is credited to Albertus Cornelis Den Brinker, Aki Sakari Harma.
Application Number | 20080215330 11/996364 |
Document ID | / |
Family ID | 37575075 |
Filed Date | 2008-09-04 |
United States Patent
Application |
20080215330 |
Kind Code |
A1 |
Harma; Aki Sakari ; et
al. |
September 4, 2008 |
Audio Signal Modification
Abstract
A method of modifying an audio signal comprises the steps of
analyzing the input audio signal (x) so as to produce a set of
filter parameters (p) and a residual signal (r), modifying the set
of filter parameters (p) so as to produce a modified set of filter
parameters (p'), and synthesizing an output audio signal (y) using
the modified set of filter parameters (p') and the residual signal
(r). The set of filter parameters (p) comprises poles
(.lamda..sub.A) and coefficients (a; c). The step of modifying the
filter parameters (p) involves interpolating lattice filter
reflection coefficients (c) so as to scale the spectral envelope of
the audio signal.
Inventors: |
Harma; Aki Sakari;
(Eindhoven, NL) ; Den Brinker; Albertus Cornelis;
(Eindhoven, NL) |
Correspondence
Address: |
PHILIPS INTELLECTUAL PROPERTY & STANDARDS
P.O. BOX 3001
BRIARCLIFF MANOR
NY
10510
US
|
Assignee: |
KONINKLIJKE PHILIPS ELECTRONICS,
N.V.
EINDHOVEN
NL
|
Family ID: |
37575075 |
Appl. No.: |
11/996364 |
Filed: |
July 18, 2006 |
PCT Filed: |
July 18, 2006 |
PCT NO: |
PCT/IB06/52450 |
371 Date: |
January 22, 2008 |
Current U.S.
Class: |
704/265 ;
704/258; 704/E13.001; 704/E21.001 |
Current CPC
Class: |
G10L 21/00 20130101;
G10L 2021/0135 20130101; G10L 19/06 20130101; H04R 25/353
20130101 |
Class at
Publication: |
704/265 ;
704/258; 704/E13.001 |
International
Class: |
G10L 13/00 20060101
G10L013/00 |
Foreign Application Data
Date |
Code |
Application Number |
Jul 21, 2005 |
EP |
05106686.8 |
Oct 5, 2005 |
EP |
05109221.1 |
Claims
1. A method of modifying an audio signal, the method comprising:
analyzing the audio signal to produce a set of filter parameters
and a residual signal, the set of filter parameters comprising
coefficients, modifying one or more of the filter parameters to
produce a modified set of filter parameters, and synthesizing a
modified audio signal using the modified set of filter parameters
and the residual signal, wherein lattice filter reflection
coefficients are interpolated to scale an envelope of the audio
signal.
2. The method according to claim 1, further comprising producing
lattice filter reflection coefficients.
3. The method according to claim 1, further comprising using
modified lattice filter reflection coefficients.
4. (canceled)
5. A method of modifying an audio signal, the method comprising:
analyzing the audio signal to produce a set of filter parameters
and a residual signal, the set of filter parameters comprising
coefficients, modifying one or more of the filter parameters to
produce a modified set of filter parameters, and synthesizing a
modified audio signal using the modified set of filter parameters
and the residual signal, wherein poles are modified to warp a
spectral envelope of the audio signal.
6. (canceled)
7. The method according to claim 1, further comprising modifying
the frequency and/or the phase of the residual signal.
8. (canceled)
9. (canceled)
10. A device for modifying an audio signal, the device comprising:
an analysis unit for analyzing the audio signal to produce a set of
filter parameters and a residual signal, the set of filter
parameters comprising coefficients (a; c), a modification unit for
modifying one or more of the filter parameters to produce a
modified set of filter parameters, and a synthesis unit for
synthesizing a modified audio signal using the modified set of
filter parameters and the residual signal, wherein the modification
unit (40) is arranged for interpolating lattice filter reflection
coefficients to scale an envelope of the audio signal.
11. The device according to claim 10, wherein analysis unit is
arranged for producing lattice filter reflection coefficients.
12. The device according to claim 10, wherein the synthesis unit
uses modified lattice filter reflection coefficients.
13. The device according to claim 10, wherein the analysis unit and
the synthesis unit comprise a lattice filter.
14. (canceled)
15. A device for modifying an audio signal, the device comprising:
an analysis unit for analyzing the audio signal to produce a set of
filter parameters and a residual signal, the set of filter
parameters comprising coefficients, a modification unit for
modifying one or more of the filter parameters to produce a
modified set of filter parameters, and a synthesis unit for
synthesizing a modified audio signal using the modified set of
filter parameters and the residual signal, wherein the modification
unit is arranged for modifying poles to warp an envelope of the
audio signal.
16. (canceled)
17. The device according to claim 15, further comprising a signal
adaptation unit for adapting the frequency and/or the phase of the
residual signal.
18. (canceled)
Description
[0001] The present invention relates to audio signal modification.
More in particular, the present invention relates to a method and a
device for the frequency axis modification of the spectral envelope
of audio signals, such as speech signals.
[0002] It is known to modify the frequency distribution of an audio
signal. In some applications, it is desired to change the frequency
scale of a signal, for example in voice modification systems. By
scaling the frequency axis, the formants of a speech signal may be
shifted so as to change the perception of the speech signal.
However, conventional scaling methods are cumbersome as they
involve many parameters which have to be set correctly to obtain
the desired result. In addition, these scaling methods typically
involve extensive computations.
[0003] In addition to (linear) scaling, the frequency axis may be
subjected to a non-linear transformation, that is, non-linear
scaling. Non-linear scaling of the frequency axis is often referred
to as (frequency) warping. Conventional warping techniques are
computationally complex.
[0004] An example of a Prior Art frequency axis modification
technique is disclosed in U.S. Pat. No. 5,930,753 (AT&T,
Potamianos). This Prior Art technique combines frequency warping
and spectral shaping in speech recognition based upon hidden Markov
models. Speech utterances are compensated by simultaneously scaling
the frequency axis and reshaping the spectral energy contour. To
optimize warping factors, computationally burdensome maximum
likelihood techniques are used.
[0005] It is an object of the present invention to overcome these
and other problems of the Prior Art and to provide a method and a
device for modifying an audio signal, in particular frequency axis
modification of the spectral envelope of an audio signal, such as a
speech signal, which are relatively simple and involve a smaller
number of control parameters.
[0006] Accordingly, the present invention provides a method of
modifying an audio signal, the method comprising the steps of:
[0007] analyzing the audio signal so as to produce a set of filter
parameters and a residual signal, the set of filter parameters
comprising poles and coefficients,
[0008] modifying one or more filter parameters so as to produce a
modified set of filter parameters, and
[0009] synthesizing a modified audio signal using the modified set
of filter parameters and the residual signal,
wherein the step of modifying one or more filter parameters
involves interpolating lattice filter reflection coefficients so as
to scale the spectral envelope of the audio signal.
[0010] By modifying lattice filter coefficients by interpolation,
as the case may be, the spectral envelope of the audio signal can
be scaled very efficiently. That is, the scaling (interpolation) of
filter coefficients in order to scale the spectral envelope of the
audio signal can be carried out with a minimal computational effort
if the filter coefficients are the coefficients of a lattice
filter, typically called reflection coefficients. The interpolation
of the lattice filter coefficients takes place over the index
number of the parameters, the index number indicating the order of
the coefficients in the filter.
[0011] It is noted that lattice filters are well known per se, but
that their very advantageous properties for scaling audio signals
have not been recognized before the present invention was made.
Lattice filters allow a simple transformation to effect a scaling
of the spectral envelope. In contrast, Prior Art methods involve
complex calculations, such as determining the autocorrelation
function of a filter, scaling the time axis of the autocorrelation
function, and deriving the modified filter parameters from the
scaled autocorrelation function. Such Prior Art methods have a high
computational complexity, while other Prior Art methods suffer from
filter instability problems.
[0012] In the method of the present invention, the step of
analyzing may produce a set of regular filter coefficients (e.g.
the coefficients of a so-called direct form filter) which are
subsequently transformed into lattice filter reflection
coefficients. In a preferred embodiment of the present invention,
however, the step of analyzing the audio signal involves producing
lattice filter reflection coefficients. That is, the reflection
coefficients are produced directly, without a prior step of
producing regular filter coefficients. The step of analyzing the
audio signal and producing a set of filter parameters and a
residual signal preferably uses a lattice filter, as this lattice
filter will be able to use the directly produced reflection
coefficients to produce the residual signal.
[0013] Similarly, it is preferred that the step of synthesizing a
modified audio signal involves using modified lattice filter
reflection coefficients. That is, the synthesis filter preferably
is a lattice filter. This avoids the intermediary step of
converting lattice filter reflection coefficients into regular
filter coefficients.
[0014] In the method of the present invention the step of modifying
one or more filter parameters may advantageously involve modifying
poles so as to warp the spectral envelope of the audio signal. In
this manner, both scaling and warping can be carried out, thus
achieving both a linear and a non-linear transformation of the
spectral envelope of the audio signal, in the direction of the
frequency axis of the spectral envelope.
[0015] The step of modifying poles so as to warp the spectral
envelope of the audio signal may also be carried out independently,
without the step of scaling the spectral envelope. Accordingly, the
present invention also provides a method of modifying an audio
signal, the method comprising the steps of:
[0016] analyzing the audio signal so as to produce a set of filter
parameters and a residual signal, the set of filter parameters
comprising poles and coefficients,
[0017] modifying one or more filter parameters so as to produce a
modified set of filter parameters, and
[0018] synthesizing a modified audio signal using the modified set
of filter parameters and the residual signal,
wherein the step of modifying one or more filter parameters
involves modifying poles so as to warp the spectral envelope of the
audio signal.
[0019] If the method of the present invention includes warping, it
is preferred that the step of modifying one or more filter
parameters involves replacing at least some poles (.lamda..sub.A)
with a modified pole (.lamda..sub.B), where the modified pole is
given by
.lamda. B = .mu. + .lamda. A 1 + .mu. .lamda. A , ##EQU00001##
and where .mu. is a warping parameter.
[0020] In addition to modifying the (spectral) envelope of the
audio signal, the residual signal may also be modified to achieve
further audio signal modifications. More in particular, the method
of the present invention may further comprise the step of modifying
the frequency and/or the phase of the residual signal.
[0021] The present invention further provides a computer program
product for carrying out the method as defined above. A computer
program product may comprise a set of computer executable
instructions stored on a data carrier, such as a CD or a DVD. The
set of computer executable instructions, which allow a programmable
computer to carry out the method as defined above, may also be
available for downloading from a remote server, for example via the
Internet.
[0022] The invention may be implemented in software, as mentioned
above, or in hardware. Suitable hardware embodiments may include an
Application-Specific Integrated Circuit (ASIC), or a programmable
logic circuit, such as a Field Programmable Gate Array (FPGA).
[0023] The present invention additionally provides a device for
modifying an audio signal, the device comprising:
[0024] an analysis unit for analyzing the audio signal so as to
produce a set of filter parameters and a residual signal, the set
of filter parameters comprising poles and coefficients,
[0025] a modification unit for modifying one or more filter
parameters so as to produce a modified set of filter parameters,
and
[0026] a synthesis unit for synthesizing a modified audio signal
using the modified set of filter parameters and the residual
signal,
wherein the modification unit is arranged for interpolating lattice
filter reflection coefficients so as to scale the envelope of the
audio signal.
[0027] In the device of the present invention, the analysis unit is
preferably arranged for producing lattice filter reflection
coefficients. Accordingly, the analysis filter may comprise a
lattice filter, or may comprise a regular (e.g. tapped line) filter
and a conversion unit for converting regular filter coefficients
into lattice filter reflection coefficients. In alternative
embodiment, however, such a conversion unit may be included in the
modification unit.
[0028] Advantageously, the synthesis unit may use modified lattice
filter reflection coefficients. In a preferred embodiment, both the
analysis unit and the synthesis unit comprises a lattice filter. In
this embodiment, no conversion from regular coefficients into
reflection coefficients is necessary and the advantageous
properties of lattice filters are fully utilized.
[0029] In an advantageous further embodiment of the present
invention, the modification unit is arranged for modifying poles so
as to warp the spectral envelope of the audio signal. Warping
involves a non-linear transformation of the spectral envelope along
its frequency axis, which transformation allows frequency spectrum
modifications which cannot be achieved by (linear) scaling
alone.
[0030] The modification unit may arranged for modifying poles
without being arranged for interpolating lattice filter reflection
coefficients. Accordingly, the present invention also provides a
device for modifying an audio signal, the device comprising:
[0031] an analysis unit for analyzing the audio signal so as to
produce a set of filter parameters and a residual signal, the set
of filter parameters comprising poles and coefficients,
[0032] a modification unit for modifying one or more filter
parameters so as to produce a modified set of filter parameters,
and
[0033] a synthesis unit for synthesizing a modified audio signal
using the modified set of filter parameters and the residual
signal,
wherein the modification unit is arranged for modifying poles so as
to warp the envelope of the audio signal.
[0034] If the device of the present invention provides warping, the
modification unit is preferably arranged for replacing at least
some poles (.lamda..sub.A) with a modified pole (.lamda..sub.B),
where the modified pole is given by
.lamda. B = .mu. + .lamda. A 1 + .mu. .lamda. A , ##EQU00002##
and where .mu. is a warping parameter. It is noted that this
warping procedure may also carried out by a device which provides
no scaling, and that warping and scaling may be carried out
independently.
[0035] In an advantageous further embodiment, the device of the
present invention further comprises a signal adaptation unit for
adapting the frequency and/or the phase of the residual signal. In
this way, the pitch of the audio signal may be changed.
[0036] The present invention further provides a consumer device and
an audio system comprising a device as defined above. A consumer
device according to the present invention may be a mobile telephone
device, a hearing aid, an electronic game and/or game console, a
personal computer, a karaoke device, or another type of consumer
device involving audio signals, in particular speech and/or voice
signals. In addition, the present invention provides a set of
filter parameters modified by the method or device defined above,
and an audio signal modified by the method or device defined
above.
[0037] The present invention will further be explained below with
reference to exemplary embodiments illustrated in the accompanying
drawings, in which:
[0038] FIG. 1 schematically shows a parametric audio signal
modification system according to the present invention.
[0039] FIG. 2 schematically shows a first embodiment of a linear
prediction analysis filter for use in the present invention.
[0040] FIG. 3 schematically shows a first embodiment of a linear
prediction synthesis filter for use in the present invention.
[0041] FIGS. 4a & 4b schematically show a second embodiment of
a linear prediction analysis filter for use in the present
invention.
[0042] FIGS. 5a & 5b schematically show a second embodiment of
a linear prediction synthesis filter for use in the present
invention.
[0043] FIGS. 6 & 7 illustrate the scaling of lattice filter
reflection coefficients according to the present invention.
[0044] FIGS. 8 & 9 illustrate the scaling of the signal
frequency spectrum according to the present invention.
[0045] The parametric audio signal modification system 1 shown
merely by way of non-limiting example in FIG. 1 comprises a linear
prediction analysis (LPA) unit 10, a signal adaptation (SA) unit
20, a linear prediction synthesis (LPS) unit 30 and a modification
(Mod) unit 40. The signal adaptation unit 20 is optional and may be
deleted if no adaptation of the residual signal corresponding with
the audio signal is desired.
[0046] The structure of the parametric audio signal modification
system 1 is known per se, however, in the system 1 illustrated in
FIG. 1 the modification unit 40 has a novel function which will
later be explained in more detail. In addition, the linear
prediction analysis (LPA) unit 10 and the linear prediction
synthesis (LPS) unit 30 preferably have a particular design which
later will be explained in more detail with reference to FIGS. 4
and 5.
[0047] The system 1 of FIG. 1 receives an audio signal x, which may
for example be a voice (speech) signal or a music signal, and
outputs a modified audio signal y. The signal x is input to the
linear prediction analysis (LPA) unit 10 which converts the signal
into a sequence of (time-varying) prediction parameters p and a
residual signal r. To this end, the linear prediction analysis unit
10 comprises a suitable linear prediction analysis filter or its
equivalent. The prediction parameters p produced by the unit 10 are
filter parameters which allow a suitable filter, in the example
shown a linear prediction synthesis (LPS) filter contained in the
linear prediction synthesis unit 30, to substantially reproduce the
signal x in response to a suitable excitation signal. The residual
signal r (or, after any pitch adaptation or other adaptation, the
modified residual signal r') serves here as the excitation
signal.
[0048] The optional signal adaptation (SA) unit 20 allows for
example the pitch (dominant frequency) of the audio signal x to be
modified by modifying the residual signal r and producing a
modified residual signal r'. Other parameters of the signal x may
be modified using the further modification unit 40 which is
arranged for modifying the prediction parameters p and producing
modified prediction parameters p'. In the present invention, the
signal adaptation (SA) unit 20 is not essential and may be omitted,
in which case the modified (or adapted) residual signal r' would be
identical to the (original) residual signal r.
[0049] An example of a linear prediction analysis filter 10 is
illustrated in FIG. 2. The exemplary filter 10 of FIG. 2 comprises
filter units 11, weighting units 12, a control unit 13 and a
combination unit 14. The input signal x is fed to both the control
unit 13 and the first weighting unit 12. Each weighting unit 12
effectively multiplies the signal with its respective weight
a.sub.0, a.sub.1, . . . a.sub.k and outputs a weighted signal which
is fed to the combination unit 14. In the embodiment shown, the
combination unit 14 adds its input signals to produce a combined
output signal r. The weights a.sub.i (i=0 . . . k) are determined
by the control unit 13.
[0050] For speech (voice) applications, the filter 10 is preferably
designed in such a way that it models the vocal tract, the output
signal r resembling a vocal excitation signal which, when input to
the vocal tract, produces a speech signal corresponding with the
filter input signal x.
[0051] In the example of FIG. 2, each filter unit 11 has an
all-pass transfer function A(z.sup.-1, .lamda..sub.A):
A ( z - 1 , .lamda. A ) = - .lamda. A + z - 1 1 - .lamda. A z - 1 (
1 ) ##EQU00003##
with z.sup.-1 representing a unit delay and .lamda..sub.A being a
transfer function parameter defining a pole of the filter. The pole
.lamda..sub.A may be determined by the control unit 13, or may be
predetermined.
[0052] The control unit 13 determines the coefficients a.sub.i and
the pole .lamda..sub.A in such a way that these parameters define
the spectral envelope of the signal x, the residual signal r having
a substantially "flat" (that is, constant) envelope. The
coefficients a.sub.i and the pole .lamda..sub.A together form a set
of parameters which is denoted p in FIG. 1. It is noted that a
different set of parameters p may be produced for each signal time
segment, for example for each frame.
[0053] The parameters a.sub.i (i=0 . . . k) and .lamda..sub.A of
the filter 10 are fed to the modification unit 40 (FIG. 1) where
they are modified. The modified parameters are output as parameters
b.sub.i (i=0 . . . k) and .lamda..sub.B. The connections between
the weighting units 12 and the modification unit 40 are not shown
in FIG. 2 for the sake of clarity of the illustration.
[0054] It is noted that all signals are discrete time signals and
could be written as x(n), y(n) and r(n) with n being the sample
number. For the sake of brevity, however, these signals are denoted
x, y and r respectively.
[0055] The parameters b.sub.i (i=0 . . . k) of the linear
prediction synthesis (LPS) filter 30 of FIG. 3 are also used as
weighting coefficients. The filter 30 comprises filter units 31,
weighting units 32 and 32', and a combination unit 34. The
weighting units 32 each have a parameter b.sub.i (i=1 . . . k),
while the weighting unit 32' has a parameter b.sub.0.sup.-1. Those
skilled in the art will understand that for b.sub.0=a.sub.0,
b.sub.i=-a.sub.i/b.sub.0 (for i=1 . . . m) and
.lamda..sub.B=.lamda..sub.A, the synthesis filter 30 is the exact
inverse of the analysis filter 10. It is noted that m may be
different from k, in other words, the number of weighting units 32
and 32' in the synthesis filter 30 is not necessarily equal to the
number of weighting units 12 in the analysis filter 10.
[0056] The filter 30 receives a parameter set p' from the
modification unit 40 (see FIG. 1). The connections between the
elements 31, 32 and 32' of filter 30 and the modification unit 40
are not shown for the sake of clarity. The parameter set p'
comprises the coefficients b.sub.i and the pole .lamda..sub.B.
[0057] The combination unit 34, which is arranged for adding its
input signals, receives the signal r produced by the filter 10 of
FIG. 2 (it is noted that the signal r may be modified by a pitch
adaptation unit 20 as illustrated in FIG. 1, in which case the
combination unit 34 receives a signal r') and the weighted filter
signals produced by the weighting units 32. The combined output
signal of the unit 34 is fed to the weighting unit 32' having the
weight (coefficient) b.sub.0.sup.-1. The output signal of the
weighting unit 32' is the filter output signal y.
[0058] In the example of FIG. 3, each filter unit 31 has a transfer
function B(z.sup.-1, .lamda..sub.B):
B ( z - 1 , .lamda. B ) = - .lamda. B + z - 1 1 - .lamda. B z - 1 (
2 ) ##EQU00004##
with z.sup.-1 representing a unit delay and .lamda..sub.B being a
transfer function parameter or pole. The parameter .lamda..sub.B is
a modified version of the corresponding parameter .lamda..sub.A of
the filter 10 of FIG. 2, the modification resulting in a non-linear
scaling (that is, a warping) of the spectral envelope of the signal
y relative to the input signal x.
[0059] The modification of the signal parameters is carried out as
follows. Assume that a scaling of the frequency axis is required of
32/24. Accordingly, the scaling factor X equals 32/24=1.33 (it will
be understood that a scaling factor .beta. equal to 1 amounts to no
scaling).
[0060] An autocorrelation function can be determined from the
impulse response of the synthesis filter. This autocorrelation
function can be re-sampled. From the re-sampled autocorrelation
function, the new coefficients of the synthesis filter can be
determined using techniques which are well known to those skilled
in the art. Typically, this is achieved by solving the normal
equations associated with the linear predictor involved. However,
solving these equations may require extensive calculations. By way
of alternative, therefore, the present invention proposes to modify
the filter coefficients, in particular the reflection coefficients
associated with these filter coefficients.
[0061] The present inventors have found that lattice filters are
particularly suitable for implementing the present invention as the
reflection coefficients are directly available in lattice filters.
This eliminates the need of converting the regular filter
coefficients a.sub.i into reflection coefficients, and the
conversion of the modified reflection coefficients into the
modified regular filter coefficients b.sub.i.
[0062] A lattice filter embodiment of a linear prediction analysis
(LPA) filter (10 in FIG. 2) is schematically illustrated in FIG.
4a.
[0063] The filter 10' comprises filter units 11, weighting units 12
and 12', a control unit 13 and combination units 14 and 15. The
filter units 11 each have a filter transfer function A(z.sup.-1,
.lamda..sub.A), as in the conventional filter 10 of FIG. 2. The
weighting units 12 each have an associated weights (weighting
parameters) c.sub.i (i=1 . . . N), each of which is equal to the
i.sup.th reflection coefficient. The weighting units 12 also have
weights c.sub.i. The control unit 13 derives the parameters
.lamda..sub.A and c.sub.i from the input signal x, as in the
embodiment of FIG. 2.
[0064] The weighting units 12 feed the output signals of the filter
units 11 to the combination units 14 to produce a combined output
signal r. As the filter 10' is a lattice filter, it has so-called
reflection coefficients that are constituted by the weights c.sub.i
of the weighting units 12'. These units 12' feed the input signal x
(in the first stage) or an intermediate signal (in subsequent
stages) to the combination units 15, which combine these weighted
signals with the output signal of the respective filter unit 11
before feeding this output signal to the next filter unit 11.
[0065] The filter units 11 of the filter 10' are illustrated in
more detail in FIG. 4b. The filter unit 11 is shown to comprise a
first combination unit 15' (which may be identical to the unit 15
shown in FIG. 4a or may be constituted by a separate unit), a
second combination unit 16, a delay unit 17 and weighting units 18
and 19. The weighting units 18 and 19 have weighting parameters
.lamda..sub.A and -.lamda..sub.A respectively.
[0066] The lattice filter 10' has the advantage of being eminently
suitable for scaling the spectral envelope of the input audio
signal as the (reflection) coefficient of the filter are directly
accessible.
[0067] A lattice filter embodiment of a linear prediction synthesis
(LPS) filter (30 in FIG. 3) is schematically illustrated in FIG.
5a. The lattice filter 30' comprises filter units 31, weighting
units 32, 32' and 32'', and combination units 34, 34' and 35. The
weighting units 32, 32' and 32'' each have an associated weighting
parameter d.sub.i (i=1 . . . N). The combination units 34, which
are arranged for adding its input signals, receive the signal r
produced by the filter 10 of FIG. 2 (or a corresponding pitch
modified signal r') and the weighted filter signals produced by the
weighting units 32. The combined output signal of the units 34 is
the filter output signal y.
[0068] Each filter unit 31 has a transfer function B(z.sup.-1,
.lamda..sub.B), with z.sup.-1 representing a unit delay and
.lamda..sub.B being a transfer function parameter. The parameter
(or pole) .lamda..sub.B is a modified version of the corresponding
parameter .lamda..sub.A of the filter 10 of FIG. 2, the
modification resulting in a non-linear frequency scaling (warping)
of the spectral envelope of the signal y relative to the spectral
envelope of the signal x.
[0069] The filter units 31 of the filter 30' are illustrated in
more detail in FIG. 5b. The filter unit 31 is shown to comprise a
first combination unit 35' (which may be identical to the unit 35
shown in FIG. 5a or may be constituted by a separate unit), a
second combination unit 36, a delay unit 37 and weighting units 38
and 39. The weighting units 38 and 39 have weighting parameters
.lamda..sub.B and -.lamda..sub.B respectively.
[0070] A (linear or proportional) scaling of the spectral envelope
can be achieved by a suitable transformation of the parameters.
More in particular, a frequency mapping may be achieved according
to the formula:
f'=.beta.f.sub.s (3)
where f' is the modified frequency, .beta. is a scaling factor and
f is the original frequency. Any modified frequency values may be
determined by scaling the (reflection) coefficients of the filters
along their axis using the same scaling factor .beta..
[0071] For example, if the frequency axis is to be scaled by a
scaling factor of 0.5 (that is, .beta.=0.5), then the filter
coefficients are scaled using this scaling factor 0.5. The new
1.sup.st coefficient, for example, obtains the value of the
original 2.sup.nd coefficient, while the new 2.sup.nd coefficient
obtains the value of the original 4.sup.th coefficient. In this
example, the number of coefficients is also halved.
[0072] For other values of .beta., for example .beta.=0.3 or
.beta.=2.0, coefficients take on values from intermediate
positions. When .beta.=0.3, for example, new coefficient no. 3
takes on the value of old coefficient no. 10 (10.times.0.3=3) but
new coefficient no. 2 assumes the value corresponding with
(non-existent) original coefficient no. 6.667. These intermediate
values are determined using interpolation techniques known per se,
such as Lagrange interpolation. This will later be illustrated with
reference to FIGS. 6 and 7.
[0073] A non-linear scaling or warping of the spectral envelope can
be achieved by a suitable transformation of the parameters. More in
particular, a frequency mapping may be achieved that can be
described by the formula:
.theta. ' = .theta. + 2 arctan ( .mu. sin ( .theta. ) 1 - .mu. cos
( .theta. ) ) , ( 4 ) ##EQU00005##
where .theta. is the frequency, normalized with respect to the
sampling frequency f.sub.s:
.theta.=2.pi.f/f.sub.s. (5)
[0074] This frequency mapping (that is, non-linear scaling of the
frequency axis) is obtained when the filter parameters
.lamda..sub.A are transformed according to:
.lamda. B = .mu. + .lamda. A 1 + .mu. .lamda. A ( 6 )
##EQU00006##
where .mu. is the warping parameter with -1<.mu.<1. It can be
seen that for .mu.=0, no warping occurs as
.lamda..sub.B=.lamda..sub.A. Using formulae (3), (4) and (5), a
desired linear and/or non-linear scaling of the frequency axis can
be obtained for given values of .beta. and .mu..
[0075] From formula (6) it is clear that linear prediction
synthesis filters based on all-pass sections, such as the filters
30 and 30', are advantageous as the filters always have the same
structure, regardless of the chosen warping factor. Only the
parameter .lamda..sub.B of the all-pass sections changes as a
function of the warping parameter .mu..
[0076] The effects of scaling are illustrated in FIGS. 6-9. FIG. 6
shows exemplary reflection coefficient values (RCV) as a function
of the coefficient index (CI) denoted i in FIGS. 4a and 5a. The
reflection coefficient values of FIG. 6 represent the coefficients
d.sub.i of the filter 30' shown in FIG. 5a in the absence of
scaling: the scaling factor .beta. equals 1 and d.sub.i=c.sub.i for
all values of i. FIG. 7 shows the same coefficients when scaled
with a scaling factor .beta. equal to 32/24=1.333. It can be seen
that the original coefficient values have been redistributed, thus
creating a new set of coefficients. For example, the value of
original coefficient no. 12 has been assigned to new coefficient
no. 16 (as 16=12.times.32/24), while new coefficient no. 15 has
received the interpolated value corresponding with non-existent
original coefficient no. 11.25 (as 15=11.25.times.32/24). In
addition, the number of coefficients has increased form 24 to
32.
[0077] In FIG. 8, the magnitude (M) of the amplitude spectrum of
the synthesis filter is shown, in decibels (dB), as a function of
the frequency (f) in the absence of scaling: .beta.=1. After
scaling with a scaling factor .beta.=32/24, the frequency spectrum
has been compressed, the peak previously located around 2.5 kHz (P)
now being located around 1.9 kHz (P'), and the peak originally
located at approximately 6.5 kHz (Q) now being located around 5.0
kHz (Q'), as illustrated in FIGS. 8 & 9. It can therefore be
seen that the present invention allows a very effective scaling of
the spectral envelope of audio signals.
[0078] It is noted that the merely exemplary spectral envelope of
FIG. 8 has been extrapolated to produce the spectral envelope of
FIG. 9. This extrapolation of the spectral envelope is the result
of the scaling factor .beta. being larger than 1 and is achieved
without extrapolating the coefficients (FIGS. 6 & 7). Instead,
some coefficient values are the result of an interpolation.
[0079] The present invention is based upon the insight that linear
and non-linear scaling operations of an audio signal, such as a
speech signal, can be effected by modifying only two control
parameters. The present invention benefits from the further
insights that the reflection coefficients of lattice filters are
particularly suitable for audio signal scaling, and that warping
may be carried out effectively using a synthesis filter based on
all-pass sections.
[0080] It is noted that any terms used in this document should not
be construed so as to limit the scope of the present invention. In
particular, the words "comprise(s)" and "comprising" are not meant
to exclude any elements not specifically stated. Single (circuit)
elements may be substituted with multiple (circuit) elements or
with their equivalents.
[0081] It will be understood by those skilled in the art that the
present invention is not limited to the embodiments illustrated
above and that many modifications and additions may be made without
departing from the scope of the invention as defined in the
appending claims.
* * * * *